Methodology

Normalization

Each metric's raw value maps to a 0–100 score against the min and max of published banks in the same cohort — not a hardcoded target. The mapping depends on the metric's normalization strategy:

Linear

Straight min–max scaling. A value at the cohort minimum scores 0, the maximum scores 100. Flipped when lower is better.

Log-scaled

Same min–max scaling, applied after a log(1+x) curve. Used for metrics that span orders of magnitude (context window, max output tokens) so a jump from 4K→128K tokens doesn't dwarf 128K→256K.

Inverse

A 1/x curve, min–max scaled. Used for price metrics — it already rewards cheap values, so the metric's "lower is better" flag isn't needed to flip it.

Pass-through

The raw value is clamped into 0–1 as-is, for metrics already expressed on that scale. Boolean metrics (does it support tool calling?) skip normalization entirely: true/false maps straight to 100/0, flipped when lower is better.

Weights

Each metric's normalized score is multiplied by a weight and summed. This table is read live from the workspace's scoring config — if a weight changes, this table changes with it, on the next request. In AI Model Rankings, benchmark scores and pricing carry the most weight because they are what the ranking claims to be about; capability booleans carry the least, because they gate what a model can do rather than how well it does it.

Missing data

Not every bank has a verified value for every metric. One of three policies decides what happens — this vertical currently runs exclude:

penalize

Missing metrics count as 0 against the full weight total. There's no way to improve a score by leaving a weak metric undocumented — an unverified value costs more than a disclosed bad one.

exclude

Missing metrics are dropped entirely and the remaining weights are renormalized to 100%.

insufficient-data

Behaves like exclude, but returns no score at all when less than half of the total weight has data — better unranked than misleading.

Confidence adjustment

Every metric value is tagged with a confidence level when it's recorded. The base score is multiplied by the average of the confidence factors across every metric that had data:

Enabled for this vertical: true.

Freshness adjustment

Data goes stale. Once the oldest verified metric value on an item passes 120 days, the score decays linearly — down to a 15% penalty once that data reaches 240 days old, where it bottoms out. One old metric drags down the whole item's score even if everything else was just verified — that's deliberate: it's the stalest data point that determines whether the score can be trusted.

In AI Model Rankings, prices are re-verified daily against three independent sources and models are canary-called with a real request — so a frontier model's score only decays if its lab's own published numbers stop being checkable.

Enabled for this vertical: true.

What we don't do

• — Affiliate and sponsored relationships never influence a score. The scoring pipeline above has no input for "paid," "sponsored," or "featured." This vertical doesn't run affiliate links at all.
• — No manual override on a published score. The only way to change a ranking is to change the underlying metric data, or the methodology itself — which gets a new version number.
• — Agent- and import-suggested data never publishes directly. It queues for human review first. Every public fact requires a cited source.

Reproduce this score

Pick any published bank below. The breakdown is recomputed right now, from the live database, using the exact pipeline documented on this page — the same one the item page's score breakdown and the ranking job both use.

Every comparison, its own methodology

The engine develops a rating methodology with each workspace — same math, that workspace's own metrics, weights, and policies. Open any of them: